Ijcot v4p303

International Journal of Computer & Organization Trends – Volume 4 – January 2014

Image Registration Methods and Validation Techniques Mr P Subramanian1, Ms R Indumathi2 1

Associate Professor, ECE Department, Aarupadai Veedu Institute of Technology, Vinayaka Missions University, 2 Lecturer, ECE Department, Aarupadai Veedu Institute of Technology, Vinayaka Missions University,

Abstract - Image registration is the process of aligning two or more images into a common coordinate system. The images can be of the same object taken at different time instants or angles or by different sensors. The images can also be of different objects. The applications of image registration include medical image analysis, neuroscience, computer vision, astrophysics, military applications etc. Many methods are currently available for image registration. It is also necessary to estimate the accuracy of the image registration process. In this paper we review the different image registration methods as well as the validation techniques.

Keywords – image registration, validation techniques, Image alignment

resolution, the problem is often cast into the continuous framework. The images are considered as functions of real arguments, the image coordinates. The correspondence between the discrete and continuous versions of the image is established using interpolation. The crudest method is the nearest-neighbor, and the most often used one is linear interpolation. Among the high-end methods, spline interpolation [3] provides the best tradeoff between accuracy and the computational cost. Occasionally, the image model occupies more dimensions than the original data. The main advantage of this approach is a more global vision of the algorithm, which increases its robustness.

I. INTRODUCTION

3.1.2. Transform based Image Registration.

Transform-based algorithms exploit properties of the Fourier, Wavelet, Hadamard and other transforms, making use of the fact that certain deformations manifest themselves more clearly in the transform domain. These methods are used mainly in connection with linear deformation fields. Nevertheless, there are examples of methods that estimate locally linear optical flow using Gabor filters [4] and B-spline wavelets [5]. Typical characteristics of the transforms employed are linearity and independence on the actual image contents.

Image registration typically involves the following steps namely Feature detection, Feature matching, transform model estimation and Image re sampling & transformation. In feature matching, the salient and distinctive objects in both reference and sensed images are detected. These include the closed boundary regions, edges, contours, line intersections and corners. The correspondence between the features is established in the next step. Image re sampling, consists of transforming the sensed image using the optimal parameters found in the previous step. The choice of the appropriate type of resampling technique depends on the trade-off between the demanded accuracy of the interpolation and the computational complexity.

II. CLASSIFICATION Image registration methods can be classified based on the feature space employed and the warp space employed.

3.1. Based on Feature Space.

Based on the feature space employed, the image registration algorithms can be classified as pixel-based, transform-based and feature-based.

3.1.1. Pixel based Image Registration.

Pixel-based algorithms work directly with the pixel values of the images being registered. Preprocessing is often used to suppress the adverse effects of noise and differences in acquisition [1] or to increase pixel resolution [2]. It is possible to work directly with the pixel values on the discrete coordinate grid. However, to get a sub pixel

ISSN: 2249-2593

3.1.3. Feature based Image Registration.

Feature-based algorithms work on a set of characteristic features extracted from the images. The dimensionality of the features is usually drastically smaller than the dimensionality of the original image data. The extraction process is highly non-linear, mostly using thresholding. Landmark based methods [6-9] use a relatively small and sparse set of landmarks. These are important points which can be (manually or automatically) identified in both images. Extrinsic markers refer to specifically designed artificial features attached to the object (or subject, in medical imaging) before acquisition to serve as landmarks. Unfortunately, extrinsic markers are difficult to deploy. In medical imaging they are not patient friendly either. If extrinsic markers are not available, we have to content ourselves with features intrinsic to the images. In that case, however, the automatic landmark identification suffers from lack of robustness. The manual landmark identification is often tedious, time-consuming, imprecise, and unreproducible. If the images cannot be characterized using points, it might be more appropriate to use curves such as edges [10] or volume boundaries [11].

Likewise, in the case of 3D data, surfaces can be used

http://www.ijcotjournal.org

Page 11

International Journal of Computer & Organization Trends – Volume 4 – January 2014 instead of working with the complete volumes. Popular features are also templates, small sub-images of important regions [12], which can be used directly, or which can form a higher level feature map [13].

3.2. Search Space based Classification.

One of the important factors to categorize registration algorithms is the search space used. We also call it a warp space, because it contains warping functions. Warping functions are candidate solutions of the registration problem. From the analogy between warping and deformation, the deformation (warping) functions play also the role of correspondence functions. Because we work with finite memory computers, every warping function from the search space is described by a finite set of real parameters (from a set of permissible values) by means of a warping model. We classify the warping models according to the number of parameters and the spatial extent of the area influenced by a single parameter.

3.2.1. Local Models.

The deformation function sought after is basically unconstrained, or belongs to a very large and unrestrictive functional space. We seek the values of this deformation at a very fine grid, usually coinciding with pixel locations.

These methods are formulated either as variational, defining a scalar criterion to minimize, or more generally using partial differential equations (PDE). The continuously defined deformation function minimizes a given criterion, or solves a given PDE. The essence of these methods is thus entirely in the criterion. The PDE come from the optical flow approach (gradient methods) [14], viscous fluid model [15-17], and elastic deformations with physical analogs or without it. Sometimes the deformation function is also modeled indirectly. For example, it can be modeled using a potential field [18].

This reduces the dimensionality of the problem, at the expense of reduced generality of the deformation.

Discretization allows for integer only displacements at pixel points [19].

3.2.2. Global Models.

Global methods describe the correspondence function using a global model with a relatively small number of parameters [20]. The model mostly consists of expressing the warping function in a linear [21], global polynomial [22] or harmonic basis [23,24]. For these methods, the deformation model corresponding to a specific warp space is as important as the criterion being minimized.

3.2.3. Semi Local Method In between the two extremes are semi-local models, using a moderate number of parameters with local influence. A grid of control points is placed over the image. Their spacing corresponds loosely to knot or landmark density. By changing the spacing, we can approach either of the limit cases or choose a compromise offering the best tradeoff. Such models were used in the

ISSN: 2249-2593

context of motion estimation [25]. B-spline models have also been independently used [26].

3.2.4. Image Dependent Models.

It is sometimes useful to adapt the warping model to the images considered. Quadtree based deformation model [27] is refined only where it is needed. In feature based methods, the basis functions of the warping model can be placed where the features are. The deformation field is interpolated in regions where no information is available.

Typical examples are radial basis functions such as thin plate splines [28].

3.2.5. Elastic Registration In this method, introduced by Bajcsy [29] et al the estimation of the geometric deformation is reduced to the search for the ‘best’ parameters. Here the images are viewed as pieces of a rubber sheet, on which external forces stretching the image and internal forces defined by stiffness or smoothness constraints are applied to bring them into alignment with the minimal amount of bending and stretching. The feature matching and mapping function design steps of the registration are done simultaneously.

Fluid registration methods make use of the viscous fluid model to control the image transformation. The reference image is here modeled as a thick fluid that flows out to match the sensed image under the control of the derivative of a Gaussian sensor model. This approach is mainly used in medical applications. Other examples of non-rigid methods are diffusion based registration, level sets registration, and optical flow based registration.

III. VALIDATION TECHNIQUES One of the challenges in the development of image registration algorithms is their validation. It is essential to provide an estimate of the accuracy of the image registration. Errors can occur in the registration process in each of its stages. There might be physical differences in the images which might be hard to distinguish from registration inaccuracies. In this section the basic error classes and validation strategies are reviewed.

The errors that can occur in the image registration can be brought under the classes of localization error, matching error and alignment error. Localization error is due to the displacement of the CP coordinates due to their inaccurate detection. Though it cannot be measured directly for a given image, the expected localization error can be estimated from the computer simulation studies of various image types. Localization error can be reduced by selecting an optimal feature detection algorithm for the given data but usually there is a tradeoff between the number of detected CP candidates and the mean localization error. Matching error is measured by the number of false matches when establishing the correspondence between CP candidates. This error often leads to failure of the registration process and can be avoided by the use of robust matching algorithms.

http://www.ijcotjournal.org

Page 12

International Journal of Computer & Organization Trends – Volume 4 – January 2014 Alignment error denotes the difference between the mapping model used for the registration and the actual between-image geometric distortion.

In many applications, the only practical approach for estimation of the registration accuracy is to visually inspect the images. Observers look at the registered images, using contour overlays for inter-modality registration [30] or difference images for intra-modality registration [31-36] and then qualitatively classify the registration solution as good, average or poor. This is a subjective and qualitative measure and it does not provide true displacement vectors between the two images.

A displacement field is generated, referred to as true displacement field, and is applied to an image to obtain a deformed image. The image registration algorithm can be run using the original and the deformed image, and the computed displacements can be compared to the true ones.

IV. CONCLUSION In this paper we have discussed the various image registration techniques, classes of errors that can occur in the image registration process and validation strategies for the registration process. Though many techniques are available there is no hard and fast rule for selecting a technique for a particular application. Also the automation of image registration process is also under research. New methods like polar transforms, sub pixel, swarm optimization, ant colony algorithms have been proposed. A combination of the methods can also be considered for specific applications.

REFERENCES [1] Jan Kybic, Philippe Th_evenaz, and Michael Unser, \Multiresolution spline warping for EPI registration," in Proceedings of SPIE, Denver, Colorado, July 1999, vol. 3813, pp. 571-579, SPIE. [2] Ruzena Bajcsy and Stane Kovafcifc, “Multiresolution elastic matching," Computer Vision, Graphics, and Image Processing, vol. 46, pp. 1-21, 1989.

[3] M. Unser, A. Aldroubi, and M. Eden, B-spline signal processing: Part I Theory," IEEE Transactions on Signal Processing, vol. 41, no. 2, pp.

821{832, Feb. 1993.

[4] A. Spinei, D. Pellerin, and J. H_erault, “Spatiotemporal energy-based method for velocity estimation," Signal Processing, vol. 65, pp. 347{362, 1998.

[5] Y.Wu, T. Kanade, J. Cohn, and C. Li, “Image registration using wavelet-based motion model," in Proceedings of Image Registration Workshop, Jacqueline Le Moigne, Ed., 1997.

[6] F. Bookstein, Morphometric Tools for Landmark Data: Geometry and Biology, Cambridge University Press, 1997.

[7] K. Rohr, H. S. Stiehl, R. Sprengel, W. Beil, T. M. Buzug, J. Weese, and M. H. Kuhn, “Point-based elastic registration of medical image data using approximating thin-plate splines," in Visualization in Biomedical Computing, Karl Heinz HFohne and Ron Kikinis, Eds., pp. 297-306.

Springer-Verlag, 1996.

[8] B. McGregor, “Automatic registration of images of pigmented skin lesions," Pattern Recognition, vol. 31, no. 6, pp. 805-817, 1998.

[9] Fred L. Bookstein, \Principal warps: Thin-plate splines and the decomposition of deformations," IEEE Trans. Pattern Anal. Mach. Intell., vol. 6, no. 6, pp. 567-585, June 1989.

[10] C. Davatzikos, J. Prince, and R. Bryan, “Image registration based on boundary mapping," IEEE Transactions on Medical Imaging, vol. 15, no.

1, Feb. 1996.

ISSN: 2249-2593

[11] C. Huang, W. Chang, L. Wu, and J. Wang, “Three-dimensional PET emission scan registration and transmission scan synthesis," IEEE Transactions on Medical Imaging, vol. 16, no. 5, Oct. 1997.

[12] P. R osch, J.Weese, T. Netsch, M. Quist, G.P. Penney, and D.L.G. Hill, “Robust 3D deformation field estimation by template propagation," in Proceedings of MICCAI, 2000.

[13] Fai Yeung, F. Levinson, Dongshan Fu, and Kevin J. Parker, “Feature-adaptive motion tracking of ultrasound image sequences using a deformable mesh," IEEE Transactions on Medical Imaging, vol. 17, no.

6, pp. 945-956, Dec. 1998.

[14] B. Horn and B. Schunck, “Determining optical flow," Artificial Intelligence, vol. 17, pp. 185-203, 1981.

[15] Gary Christensen, Deformable Shape Models for Anatomy, Ph.D. thesis, Washington University, Saint Louis, Mississippi, 1994.

[16] Gary Christensen, S. Joshi, and M. Miller, \Volumetric transformation of brain anatomy," IEEE Transactions on Medical Imaging, vol. 16, Dec. 1997.

[17] M. Bro-Nielsen and C. Gramkow, “Fast fluid registration of medical images," in Visualization in Biomedical Computing, Karl Heinz H ohne and Ron Kikinis, Eds., pp. 267{276. Springer-Verlag, 1996.

[18] T. Schormann, S. Henn, and K. Zilles, “A new approach to fast elastic alignment with applications to human brains," in Visualization in Biomedical Computing, Karl Heinz H ohne and Ron Kikinis, Eds. Springer-Verlag, 1996.

[19] Ulrik Kjems, Stephen C. Strother, Jon Anderson, Ian Law, and Lars Kai Hansen, “Enhancing the multivariate signal of water PET studies with a new nonlinear neuroanatomical registration algorithm," IEEE Transactions on Medical Imaging, vol. 18, no. 4, pp. 306{319, Apr. 1999.

[20] J. Bergen, P. Anandan, K. Hanna, and R. Hingorani, “Hierarchical model based motion estimation," in Second European Conference on Computer Vision (ECCV'92). 1992, pp. 237{252, Springer-Verlag.

[21] Philippe Th_evenaz, Urs E. Ruttimann, and Michael Unser, “A pyramid approach to subpixel registration based on intensity," IEEE Transactions on Image Processing, vol. 7, no. 1, pp. 1{15, Jan. 1998.

[22] N. Sicotte, R. Woods, and J. Mazziotta, “Automated image registration using a 105 parameter non-linear model," Neuroimage, vol. 3, no. 3, June 1996, Second International Conference on Functional Mapping of the Human Brain.

[23] S. Kiebel, J. Ashburner, J. Poline, and K. Friston, “MRI and PET coregistration-a cross validation of statistical parametric mapping and automated image registration," Neuroimage, , no. 5, 1997.

[24] John Ashburner and Karl J. Friston, “Nonlinear spatial normalization using basis functions," in Brain Warping, Arthur W. Toga, Ed., pp.

254{266. Academic Press, San Diego, 1999.

[25] Richard Szeliski and James Coughlan, “Spline-based image registration," International Journal of Computer Vision, vol. 22, pp.

199{218, 1997.

[26] Colin Studholme, R. Todd Constable, and James S. Duncan, “Incorporating an image distortion model in non-rigid alignment of EPI with conventional MRI," in International Conference on Information Processing in Medical Imaging, Lecture Notes in Computer Science, Visegrfad, Hungary, June 1999, pp. 454{459, Springer-Verlag.

[27] Richard Szeliski and Heung-Yeung Shum, \Motion estimation with quadtree splines," IEEE Trans. Pattern Anal. Mach. Intell., vol. 18, no.

12, pp. 1199-1207, Dec. 1996.

[28] H. Lester and S.R. Arridge, “Summarising fluid registration by thinplate spline warps with many landmarks," in Medical Image Understanding and Analysis, July 1997, [29] R. Bajcsy, S. Kovacic, Multiresolution elastic matching, Computer Vision, Graphics and Image Processing 46 (1989) 1–21.

[30] Fitzpatrick, J.M., Hill, D.L.G, Shyr, Y., et al., "Visual assessment of the accuracy of retrospective registration of MR and CT images of the brain", IEEE Trans. Med. Imag,17: 571-585, 1998.

[31] Barber, D.C., Tindale, W.B., Hunt, E., et al. , "Automatic registration of SPECT images as an alternative to immobilization studies", Physics in medicine and biology, 40:449- 463, 1995.

[32] Joshi, S.C. and Miller, M.I., "Landmark Matching via Large Deformation Difoeomorphisms", IEEE Transactions on Image Processing, 9(8), pp. 1357-1370, 2000.

[33] Rabbitt, R.D., Weiss, J., Christensen, G., et al., "Mapping of hyperelastic deformable templates using the finite element method", Proc. of SPIE, International Symposium on Optical Science, Engineering, and Instrumentation, vol. 2573, pp. 252-265, 1995.

http://www.ijcotjournal.org

Page 13

International Journal of Computer & Organization Trends – Volume 4 – January 2014 [34] Thirion, J. -P., "Image matching as a diffusion process: an analogy with Maxwell's demons", Medical Image Analysis, Vol. 2, Issue: 3, pp.

243-260, 1998.

[35] Ferrand, M., Warfield, S., Nabavi, A., et al., "Registration of 3D intraoperative MR images of the brain using a finite element biomechanical model", Medical Image Computing and ComputerAssisted Intervention MICCAI'2000, Vol. 1935, pp. 19-28, 2000.

[36] Hagemann, A., Rohr, K., Stiehl, H.S., et al., "Biomechanical modeling of the human head for physically based, nonrigid image registration", IEEE Trans. on Medical Imaging, 18(10): 875-884, 1999.

[37] Rohr, K., Fornefett, M., and Stiehl, H.S., "Approximating thin-plate splines for elastic registration: integration of landmark errors and orientation attributes", In A. Kuba, M. Samal, and A. Todd-Pokropek, editors, Information Processing in Medical Imaging: Proc. 16th International Conference (IPMI'99), volume 1613 of Lecture Notes in Computer Science, pages 252-265. Springer Verlag, 1999 [38] Feature based Image Registration, Thesis by Somaraju Boda, NIIT Rorkela,

ISSN: 2249-2593

http://www.ijcotjournal.org

Page 14